Researchers propose the concept of 'bounded sovereignty' to address the challenge of controlling AI models when they are deployed through APIs or managed endpoints, where the deployer may not have full control over the model. They argue that access conditions across four layers - data, model, infrastructure, and interaction - determine which control protocols can be executed in practice. The study introduces a typology of access levels and a matrix of protocol requirements by layer, as well as the concept of 'sovereignty discount cost', which refers to the additional costs incurred when substituting for missing access through contracts or architecture. An experiment using synthetic case simulations found that complete logs improve diagnosis, pre-execution gateways enable intervention, and trace access and model-version control strengthen post-incident explanation. --- Why it matters: This research matters because it highlights the limitations of current AI control protocols, which often assume full control over the model. Engineers and researchers working on AI safety need to consider these access assumptions explicitly when designing control protocols to ensure they are effective in real-world scenarios. Source: https://arxiv.org/abs/2608.19216